Accurate enough to be useful, and never presented as more than that. Every reading
comes with a range around the number, and a Low, Moderate or Good confidence rating
that tells you how much the app had to work with. A side on shot with a gate in
frame gets a tight band. A distant photo at a bad angle with nothing of known size
in it gets a wide one, and the app says why.
We publish what we have measured. Our first test used the model we launched with:
on 70 scale-weighed cattle photographed with nothing of known size in frame, the
typical error was about 23 per cent, and the true weight fell inside the shown range
about four times in ten. Production has since moved to a newer model. On the first
ten of those same cattle it ran at a typical error of about 9 per cent, with the true
weight inside the range seven times out of ten. Ten animals is too few to promise
anything on, so we are re-running the full set and will put that figure here when
it exists.
The formula itself is within about six per cent when fed a real tape, so the error
is in reading the measurements off pixels, and that is what the ruler sheet and the
known length are for. The honest position is that it will not match a set of load
bars and is not trying to. The comparison worth making is against the guess you
would otherwise have made over the fence.